Re: USB EHCI patch for 0.14.0?

[David Ahern <daahern@cisco.com>]

On 03/09/11 14:28, Erik Rull wrote:
> David Ahern wrote:
>> On 03/09/11 08:40, erik.rull@rdsoftware.de wrote:
>>> But some things are not working and cause my Windows guest to stop
>>> booting
>>> or getting slowed down:
>>> -device usb-tablet
>>> -device usb-mouse
>>> do not really work. (I have connected a PS/2 mouse to have no
>>> interference
>>> with the rest of the USB system that works fine without the patch)
>>> If I add them to the command line, windows does not boot up (it hangs
>>> before the GUI comes up with ~ 12% CPU time on the host side)
>>> If I add them at runtime via the qemu console it has no influence to the
>>> guest - I still see no possibility grabbing the mouse to the client
>>
>> Did those work with the previous qemu-kvm releases you tested?
> 
> I've tried it long time ago with kvm-88, but there it was working but
> extremely slow. (At least the mouse was reacting there)

I thought you tried the EHCI patch against a recent qemu-kvm version --
like December 2010 or January 2011.


> Hm, okay.
> As far as I understood it, the auto-add feature should be similar to the
> USB 1.1, right? It seems to work basically but not fully - the system is
> somehow slowed down, maybe the polling timer is too fast? (I will play a
> little bit with that and review and compare as well the auto routine)
> 
> And the usb-tablet should be an uhci-emulated component, that should
> then not interfere with the ehci-emulation, right?

My proposal from July 2010 was to have emulated devices state their
version and have host devices try EHCI then UHCI. This means that the
tablet device is attached to UHCI and a host USB key is attached to
EHCI. Like this:

info usb
  Device 0.1, Port 1, Speed 12 Mb/s, Product QEMU USB Tablet
  Device 1.1, Port , Speed 480 Mb/s, Product DT 101 II
(qemu)

That's what the patch I sent does.

EHCI does have a lot higher frame rate and in its current form does have
a noticeable impact on CPU usage when devices are connected and one of
the lists is activated. If the OS deactivates the controller when there
is nothing to do (e.g., not actively talking to the device), cpu usage
goes down.

I just tried a few scenarios with the X and V versions of those GUI
based guests and an external key attached to EHCI bus worked fine and
the usb tablet also worked fine.

I did notice some differences in command syntax. For instance, my
scripts still use the older -usbdevice tablet syntax and I did not see
the USB stall message. Switching to '-device usb-tablet' did generate
the message at boot (though overall it seems to be harmless).

Nothing fancy with the setup -- ide drive, virtio or e1000 networking,
ac97 sound, no-hpet and usb tablet devices.

In the time it took to write this response about 900MB was transferred
to the usb key at about 1.7-1.8MB/sec rate.

David

> 
> If you have any additional hints where to start digging, just let me know.
> 
> Best regards,
> 
> Erik
>